基于机器视觉的行人检测技术研究

发布时间：2018-04-27 17:45

本文选题：行人检测 + 感兴趣区域提取　；参考：《长春工业大学》2017年硕士论文

【摘要】：随着智慧城市的建设需要,行人检测技术的研究与发展备受关注。行人检测技术是图像处理、模式识别、机器学习等多学科交叉的研究课题。此外,行人检测在智能汽车,安防等领域也具有广阔的应用空间。目前,行人检测技术的研究已经取得了很大的进展,但是依然存在很多难点。例如,提取的感兴趣区域窗口过多;提取的特征对行人描述不足;特征维度过高以及分类器等方面还有待进一步的提高等。为了解决上述问题,本文从感兴趣区域提取、特征提取和分类器三个方面对行人检测技术开展相关研究。本文的主要研究内容如下:(1)在感兴趣区域的研究中,由于行人检测中使用广泛的多尺度滑动窗口法产生了过多的检测窗口。本文提出了一种基于路面约束和图像分割来提取感兴趣区域的方法来减少待检测窗口的数量。首先使用FCM算法对预处理之后的灰度图像进行分割,标记处理之后的连通区域。另一方面通过区域生长、边缘检测、霍夫变换等算法对行人检测图像中的路面区域进行提取,并根据生活中行人与路面的关系,利用所提取的路面对已标记的连通区域进行筛选,符合条件的连通区域即为感兴趣区域。为了减弱光照的影响和突出行人区域,本文在预处理中加入了直方图均衡化的方法。(2)在特征提取中,针对HOG等单一的特征对行人描述不足,融合特征维度过高等问题。本文提出了一种多尺度融合的稀疏特征对行人进行描述,以此兼顾特征描述不足和特征维度过高的问题。首先提取行人检测图像的HOG特征和LBP特征,对提取的HOG和LBP特征分别进行稀疏表示,然后将两个稀疏特征进行多尺度的融合。HOG特征的提取采用经典的3780维特征,然后对其进行稀疏表示。在LBP特征的提取中,首先将64*128的样本分块然后再提取块内的LBP特征,并对块内的特征进行归一化处理,并将所有块内的LBP特征组成图像的LBP特征。通过实验研究发现,当块大小为8*16像素时,提取的LBP特征检测效果较好。在进行稀疏特征融合的过程中,比较了在不同融合尺度下的行人检测效果。最终完成多尺度稀疏特征的融合。(3)在分类器的研究中,本文首先将KNN分类器和行人检测中经常使用的支持向量机分类器作对比。并比较了未进行参数优化的RBF核函数和线性核函数的分类效果。为了进一步提高基于RBF核函数支持向量机的分类效果,本文对RBF的参数进行了两次寻优。首先利用网格搜索算法确定最优参数的范围,在此范围的基础上利用粒子群优化算法进行参数的精确优化。将优化后的RBF核函数的支持向量机与使用线性核函数的支持向量机进行对比。
[Abstract]:With the construction of intelligent city, the research and development of pedestrian detection technology has attracted much attention. Pedestrian detection technology is an interdisciplinary research topic of image processing, pattern recognition, machine learning and so on. In addition, pedestrian detection in intelligent vehicles, security and other fields also has a wide range of applications. At present, the research of pedestrian detection technology has made great progress, but there are still many difficulties. For example, there are too many windows in the extracted region of interest; the extracted features are not enough to describe the pedestrian; the feature dimension is too high and the classifier still needs to be further improved. In order to solve the above problems, this paper studies pedestrian detection technology from three aspects: region of interest extraction, feature extraction and classifier. The main contents of this paper are as follows: (1) in the study of the region of interest, there are too many detection windows due to the extensive use of multi-scale sliding window method in pedestrian detection. In this paper, a method of extracting the region of interest based on road constraints and image segmentation is proposed to reduce the number of windows to be detected. Firstly, FCM algorithm is used to segment the preprocessed gray image and mark the connected region. On the other hand, the pavement area in pedestrian detection image is extracted by region growth, edge detection, Hough transform and so on, and based on the relationship between pedestrian and pavement in daily life, Using the extracted road surface to screen the marked connected region, the connected region that meets the conditions is the region of interest. In order to reduce the influence of illumination and highlight pedestrian area, histogram equalization method is added in this paper. In feature extraction, the single feature, such as HOG, is not enough to describe pedestrians, and the feature dimension is too high. In this paper, a multi-scale fusion sparse feature is proposed to describe pedestrians, which takes into account the insufficiency of feature description and the problem of high feature dimension. Firstly, the HOG feature and LBP feature of pedestrian detection image are extracted, and the extracted HOG and LBP features are represented sparsely respectively. Then, the two sparse features are extracted by multi-scale fusion .HOG feature using the classical 3780 dimension feature. Then it is represented sparsely. In the extraction of LBP features, the samples of 64T128 are first divided into blocks, then the LBP features in the blocks are extracted, and the features in the blocks are normalized, and all the LBP features in the blocks are formed into the LBP features of the image. The experimental results show that when the block size is 816 pixels, the LBP feature detection is better. In the process of sparse feature fusion, the pedestrian detection performance at different fusion scales is compared. Finally, the fusion of multi-scale sparse features. 3) in the research of classifier, this paper first compares the KNN classifier with the support vector machine classifier, which is often used in pedestrian detection. The classification effects of RBF kernel function and linear kernel function without parameter optimization are compared. In order to further improve the classification effect of support vector machine based on RBF kernel function, the parameters of RBF are optimized twice in this paper. Firstly, the range of the optimal parameters is determined by using the grid search algorithm, and on the basis of this range, the particle swarm optimization algorithm is used to precisely optimize the parameters. The support vector machine (SVM) of the optimized RBF kernel function is compared with the support vector machine (SVM) using linear kernel function.
【学位授予单位】：长春工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】